Artificial Intelligence

[Eldad]

6 minutes ago, MungerWunger said:

We're saved!

 

https://www.theverge.com/tech/946879/microsoft-mustafa-suleyman-ai-white-collar-jobs

It was always a sales pitch. Now that the college kids boo the commencement speaker when they mention AI, the AI Corporates pretty much all shifted in unison to this statement. 
 

Read this thread about 6 months to a year ago and see who bought it hook line and sinker. Maybe some of the same ones that thought we would never emerge from Covid in early summer 2020. 

[frommi]

Have fun coding with Claude or Codex agents. When they install packages you never know what happens .
This is like Covid for coding agents.

Edited Tuesday at 04:05 PM by frommi

[MungerWunger]

https://www.bbc.com/news/articles/ckg701v1dp6o

 

Version of AI tool 'too powerful for public' released to public

 

^ hilarious title, did the world end yet?

[ratiman]

Around 130 companies were asked some version of “is AI eating your business?” and said no. Almost all of enterprise software sits here — Salesforce, ServiceNow, Workday, Adobe, SAP, Atlassian — swatting the seat-compression question away like it’s beneath them.

Now, none of this is secret. The denials are in the prepared remarks for anyone to read. And the pricing changes are in the press releases for anyone to read. The interesting thing is that they contradict each other. Almost every one of these companies is, at the same time, tearing out per-seat pricing and replacing it with usage-based pricing.

Sit with the logic. If your software is sold per employee, and AI agents are about to do those employees’ jobs, then per-seat revenue is the exact thing AI threatens. A company that genuinely believed AI was harmless would have no reason to rebuild its entire pricing model on a deadline. When a company’s words and its pricing point in opposite directions, the pricing is the more honest signal — because it costs something.

 

https://autonomousresearchcorp.com/research/ai-revenue-headwind

 

[Libs]

6 minutes ago, ratiman said:

Around 130 companies were asked some version of “is AI eating your business?” and said no. Almost all of enterprise software sits here — Salesforce, ServiceNow, Workday, Adobe, SAP, Atlassian — swatting the seat-compression question away like it’s beneath them.

Now, none of this is secret. The denials are in the prepared remarks for anyone to read. And the pricing changes are in the press releases for anyone to read. The interesting thing is that they contradict each other. Almost every one of these companies is, at the same time, tearing out per-seat pricing and replacing it with usage-based pricing.

Sit with the logic. If your software is sold per employee, and AI agents are about to do those employees’ jobs, then per-seat revenue is the exact thing AI threatens. A company that genuinely believed AI was harmless would have no reason to rebuild its entire pricing model on a deadline. When a company’s words and its pricing point in opposite directions, the pricing is the more honest signal — because it costs something.

 

https://autonomousresearchcorp.com/research/ai-revenue-headwind

 

 

Not sure I get the logic.

 

Software co X sells mission-critical, deeply embedded software to company Y. X uses seat - based pricing, totaling say $10,000 a year.

 

Company Y lays off 20% of their employees due to efficiencies gained from AI. But they still need and value X's software. 

 

Company X simply switches to a flat $10,000 fee instead of seat-based.

 

Maybe I'm missing something; I don't see the problem.

[ratiman]

I thought the whole post was interesting. As for ERP seat vs fee based pricing, I'm not sure a major change to enterprise software business model is great news. It's easy to count and verify seats which makes it a convenient way to structure pricing. The alternative is untested. 

[Castanza]

Can't speak to whether AI pulled forward this trend, but companies SaaS companies have been switching their pricing model from per seat to usage since 2018. Many of them offered the choice for some time so it may not have been as obvious that this trend had already started. I deal with quite a bit of licensing fee bullshit and this for sure predates the AI boom. Same for PaaS and IaaS. I can see the token model accelerating this though. Covid was also a big driver of the shift. 

Edited Wednesday at 02:01 AM by Castanza

[frommi]

2 hours ago, ratiman said:

Around 130 companies were asked some version of “is AI eating your business?” and said no. Almost all of enterprise software sits here — Salesforce, ServiceNow, Workday, Adobe, SAP, Atlassian — swatting the seat-compression question away like it’s beneath them.

Now, none of this is secret. The denials are in the prepared remarks for anyone to read. And the pricing changes are in the press releases for anyone to read. The interesting thing is that they contradict each other. Almost every one of these companies is, at the same time, tearing out per-seat pricing and replacing it with usage-based pricing.

Sit with the logic. If your software is sold per employee, and AI agents are about to do those employees’ jobs, then per-seat revenue is the exact thing AI threatens. A company that genuinely believed AI was harmless would have no reason to rebuild its entire pricing model on a deadline. When a company’s words and its pricing point in opposite directions, the pricing is the more honest signal — because it costs something.

 

https://autonomousresearchcorp.com/research/ai-revenue-headwind

 

Just because the companies now offer additional pricing makes them obsolete?
The whole universe of AI will soon realize that customers hate token based pricing. For the software vendors that offer this, it doesn't matter since for them its a pass-through cost. But they will not let their AI-bill for their coding agents go rampant. See Uber. Long term this pricing tier will vanish, like it vanished in mobile data usage. And customers will try to reduce this cost increase by using free models. They didn't have to because models where subsidized, but now it will happen. (and is already as 30%+ of AI traffic is already these free models)

Model Name	Type	Blended Cost (Per 1M Tokens)	June 2026 Standing / Core Niche
GPT-5.5 (OpenAI)	Frontier (Closed)	~$11.25	The Autonomy SOTA. The absolute gold standard for multi-step DevOps, complex infrastructure engineering, and terminal agent tasks.
Claude Opus 4.8 (Anthropic)	Frontier (Closed)	~$10.00	The Academic Leader. Unrivaled at deep, PhD-level multidisciplinary reasoning and complex codebase architecture.
GPT-5.4 Mini (OpenAI)	Frontier (Closed)	~$1.68	The Commercial Standard. Built to serve massive enterprise web pipelines that require closed-loop data security.
DeepSeek-V4-Pro	Open / Free Tier	~$0.90*	The Long-Context Disrupter. Offers a massive 1-million-token context window and performance that lags behind the frontier by only a few months.
Qwen 3.5 (397B) (Alibaba)	Open / Free Tier	~$0.75*	The Global Enterprise Workhorse. The most downloaded open model on Hugging Face; unbeatable multilingual support.
Kimi K2.6 (Moonshot AI)	Open / Free Tier	~$0.60*	The Math/Logic Challenger. Top open-weights score for advanced STEM deduction and code auditing.
Gemma 4 (12B/31B) (Google)	Open / Free Tier	~$0.12*	The On-Device King. Entirely encoder-free multimodal architecture optimized to run for $0 locally on basic 16GB laptops.

[frommi]

Just to underpin what i wrote, this is rent cost for NVDA chips:

The decline started exactly when MSFT introduced its new token based pricing model for Copilot.

[frommi]

Whoever still doesn't believe me

[Peregrine]

4 hours ago, frommi said:

Whoever still doesn't believe me

 

So what's the point of using AI to write all this code if it can't be used, needs to be reviewed and understood line by line and then debugged all over again? On top of that, you're paying ever escalating prices for tokens using the frontier models.

[frommi]

5 minutes ago, Peregrine said:

 

So what's the point of using AI to write all this code if it can't be used, needs to be reviewed and understood line by line and then debugged all over again? On top of that, you're paying ever escalating prices for tokens using the frontier models.

Its still faster than writing it by hand. Especially for test code its really good, because the quality is not that important.

[Peregrine]

3 minutes ago, frommi said:

Its still faster than writing it by hand. Especially for test code its really good, because the quality is not that important.

 

I'm sure it is but I wonder if it's a glorified auto-complete and whether it'll ever be reliabily used on its own.

[frommi]

29 minutes ago, Peregrine said:

 

I'm sure it is but I wonder if it's a glorified auto-complete and whether it'll ever be reliabily used on its own.

No, because it has no "real" intelligence. To me this is fake or simulated intelligence. But that doesn't mean it can't produce working software. 

This article is spot on: https://geohot.github.io/blog/jekyll/update/2026/05/24/the-eternal-sloptember.html

Edited Wednesday at 06:28 PM by frommi

[Peregrine]

54 minutes ago, frommi said:

No, because it has no "real" intelligence. To me this is fake or simulated intelligence. But that doesn't mean it can't produce working software. 

This article is spot on: https://geohot.github.io/blog/jekyll/update/2026/05/24/the-eternal-sloptember.html

 

Thanks for this. A succint description of AI's coding issues and aligned with all the data showing the proliferation of AI slop apps with 0 adoption.

[Castanza]

1 hour ago, frommi said:

No, because it has no "real" intelligence. To me this is fake or simulated intelligence. But that doesn't mean it can't produce working software. 

This article is spot on: https://geohot.github.io/blog/jekyll/update/2026/05/24/the-eternal-sloptember.html

 

Yup, so far AI is probabilistic where human coders are deterministic. That difference is a glass ceiling for how far AI generated code can go. Still a very useful tool though especially when developed for specialized tasks. It gets you closer quicker.

[Spekulatius]

20 hours ago, ratiman said:

I thought the whole post was interesting. As for ERP seat vs fee based pricing, I'm not sure a major change to enterprise software business model is great news. It's easy to count and verify seats which makes it a convenient way to structure pricing. The alternative is untested. 

The end results is more likely be usage based. Seats (or more correctly licenses)  are just a shortcut for usage anyways. I don’t see a problem either.

[vinod1]

3 hours ago, Castanza said:

 

Yup, so far AI is probabilistic where human coders are deterministic. That difference is a glass ceiling for how far AI generated code can go. Still a very useful tool though especially when developed for specialized tasks. It gets you closer quicker.

 

I think and I read this somewhere that while AI generated code is probabilistic, the generated code can still produce deterministic output. 

 

Trivial example: Use AI to generate code a calculator. The calculator is going to be deterministic in its function.

 

Vinod

[Castanza]

2 hours ago, vinod1 said:

 

I think and I read this somewhere that while AI generated code is probabilistic, the generated code can still produce deterministic output. 

 

Trivial example: Use AI to generate code a calculator. The calculator is going to be deterministic in its function.

 

Vinod

Very true! But there is still a lot more nuance. It’s possible good enough is simple good enough though. 

[gfp]

I've had a chance to try out Anthropic's "fable 5" model today and it is very very good - I've been impressed.  It is expensive to run and they are only including it in my plan until June 22nd without paying more but progress marches on and I would reiterate that if you are not using the frontier models every day and you think you know what you are talking about when you opine about AI and this bubble and where things are headed, you are missing the plot.  Use AI every day.  get good at it.  and then form an opinion that may be worth something

[Milu]

2 hours ago, gfp said:

I've had a chance to try out Anthropic's "fable 5" model today and it is very very good - I've been impressed.  It is expensive to run and they are only including it in my plan until June 22nd without paying more but progress marches on and I would reiterate that if you are not using the frontier models every day and you think you know what you are talking about when you opine about AI and this bubble and where things are headed, you are missing the plot.  Use AI every day.  get good at it.  and then form an opinion that may be worth something

Yes fully agree with this. 

[Castanza]

8 hours ago, gfp said:

I've had a chance to try out Anthropic's "fable 5" model today and it is very very good - I've been impressed.  It is expensive to run and they are only including it in my plan until June 22nd without paying more but progress marches on and I would reiterate that if you are not using the frontier models every day and you think you know what you are talking about when you opine about AI and this bubble and where things are headed, you are missing the plot.  Use AI every day.  get good at it.  and then form an opinion that may be worth something

How are you using the models? 

[Sinbius]

8 hours ago, gfp said:

I've had a chance to try out Anthropic's "fable 5" model today and it is very very good - I've been impressed.  It is expensive to run and they are only including it in my plan until June 22nd without paying more but progress marches on and I would reiterate that if you are not using the frontier models every day and you think you know what you are talking about when you opine about AI and this bubble and where things are headed, you are missing the plot.  Use AI every day.  get good at it.  and then form an opinion that may be worth something

 

80% of american startups run on chinese AI....

[Libs]

10 hours ago, gfp said:

I've had a chance to try out Anthropic's "fable 5" model today and it is very very good - I've been impressed.  It is expensive to run and they are only including it in my plan until June 22nd without paying more but progress marches on and I would reiterate that if you are not using the frontier models every day and you think you know what you are talking about when you opine about AI and this bubble and where things are headed, you are missing the plot.  Use AI every day.  get good at it.  and then form an opinion that may be worth something

 

gfp,

 

How much would you be willing to pay? I'm trying to figure out if users will ever cough up enough to cover Anthropic's (et al) costs....

Edited Thursday at 03:05 PM by Libs

[gfp]

4 hours ago, Libs said:

 

gfp,

 

How much would you be willing to pay? I'm trying to figure out if users will ever cough up enough to cover Anthropic's (et al) costs....

 

I don't know the answer to that.  I currently pay $22 per month to OpenAI, I paid $99 for a year of Gemini (limited time offer), and I paid something like $220 for year of Anthropic's Claude "Pro"

 

I think Anthropic is the most disciplined about not giving away token usage for a big loss.

 

I'm not investing in any of the model companies and it is possible that they end up a little like the Ford CEO shaking in his boots at the thought of Chinese cars entering the US auto market.

 

So far, Anthropic has done very well finding paying customers and I personally can't imagine going back to lower tier or free versions of these subscriptions.  These are a phenomenal value.